Digitized Counterdiabatic Quantum Optimization

• Digitized counterdiabatic quantum optimization is a family of methods that add counterdiabatic driving to standard gate-based techniques, boosting finite-time ground-state success.
• It applies an approximate adiabatic gauge potential to generate additional one- and two-body operators, such as Y, YZ, and ZY terms, forming both variational and nonvariational protocols.
• Empirical benchmarks show that counterdiabatic corrections can reduce circuit depth and improve performance on Ising-type problems, especially in shallow circuit regimes.

Digitized counterdiabatic quantum optimization denotes a family of gate-based quantum optimization methods that start from adiabatic quantum optimization, add counterdiabatic driving to suppress diabatic transitions, and then digitize the resulting evolution into circuits suitable for digital hardware. In this literature, the target problem is typically encoded in an Ising-type Hamiltonian, an easy transverse-field Hamiltonian provides the initial state, and an approximate adiabatic gauge potential supplies additional one- and two-body generators such as $Y$, $YZ$, and $ZY$ terms. The paradigm includes both nonvariational protocols, usually called DCQO or DCQC, and variational descendants such as digitized-counterdiabatic QAOA (DC-QAOA), hybrid DCQO, and bias-field DCQO (Hegade et al., 2022).

1. Origins and conceptual scope

An early formulation proposed digitized-counterdiabatic quantum optimization as a gate-based optimization paradigm for the general Ising spin-glass model, emphasizing that the relevant counterdiabatic terms are naturally non-stoquastic and that even a 2-local first-order construction can yield a polynomial enhancement in finite-time ground-state success probability relative to stoquastic adiabatic quantum optimization (Hegade et al., 2022). A closely related line introduced digitized-counterdiabatic QAOA by inserting a counterdiabatic unitary into each QAOA layer, thereby importing shortcut-to-adiabaticity ideas into a shallow variational circuit and reporting improved low-depth performance on Ising models, MaxCut, the Sherrington–Kirkpatrick model, and the $P$-spin model (Chandarana et al., 2021).

Within this broad framework, the literature splits into several algorithmic interpretations. One interpretation treats digitized counterdiabatic optimization as a discretized form of adiabatic or annealing dynamics with analytically specified counterdiabatic coefficients. Another treats it as a variational ansatz-design strategy, where approximate counterdiabatic operators enlarge the reachable state manifold at fixed depth, often by adding one additional unitary per layer to QAOA-like circuits (Chandarana et al., 2021). A further development replaces outer-loop variational training with measurement-driven feedback, as in bias-field DCQO, where low-energy samples are recycled into longitudinal bias fields for the next run (Cadavid et al., 2024).

The topic is therefore not synonymous with a single ansatz. It comprises schedule-based digitized annealing with counterdiabatic corrections, counterdiabatic-enhanced QAOA, compressed “CD-only” circuits in the impulse regime, and iterative feedback protocols. A plausible implication is that the common object is not a specific circuit template but a design principle: use approximate adiabatic gauge-potential structure to improve finite-time optimization on digital hardware.

2. Common mathematical formulation

A common starting point writes the adiabatic interpolation as

$$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$

or equivalently $H_{ad}(\lambda)=(1-\lambda)H_i+\lambda H_f$, where $H_i$ is typically a transverse field and $H_f$ is the Ising cost Hamiltonian (Chandarana et al., 2021). Counterdiabatic driving augments this with an adiabatic gauge-potential term,

$$H_{cd}(\lambda)=H_{ad}(\lambda)+\dot\lambda A_\lambda,$$

with $A_\lambda$ exact only in principle and therefore usually replaced by an approximation (Cadavid et al., 2024).

The standard approximation in the cited works is the nested-commutator expansion

$YZ$0

At first order, this produces tractable one- and two-body terms. For transverse-field Ising-type problems, the first-order pool repeatedly yields local $YZ$1 terms and mixed two-body strings such as $YZ$2 and $YZ$3, which then become the building blocks of the digitized circuit (Cadavid et al., 2023).

Standard QAOA is written as

$YZ$4

where $YZ$5 and $YZ$6. DC-QAOA enlarges this to

$YZ$7

by adding a CD-generated unitary in each layer; in the formulation emphasized in the original DC-QAOA paper this changes the parameter count from $YZ$8 to $YZ$9 (Chandarana et al., 2021). In nonvariational DCQO, by contrast, the circuit is obtained directly from Trotterized evolution under the approximate CD-corrected Hamiltonian, often with analytically computed coefficients and without a classical parameter-search loop (Cadavid et al., 2024).

Several later papers compress this structure further. In the impulse regime, where the CD contribution dominates, one may discard most of the adiabatic term and keep only the CD piece over the time window where $ZY$0 is large; this yields circuits generated only by $ZY$1, $ZY$2, and $ZY$3 blocks and can reduce depth sharply (Cadavid et al., 2023). This suggests that in practice the most useful information in a finite-time adiabatic path may be concentrated in a small subset of counterdiabatic directions rather than in a faithful discretization of the full schedule.

3. Algorithmic families and representative variants

The field now contains several distinct but related families.

Variant	Defining feature	Representative paper
DC-QAOA	Adds one CD unitary per QAOA layer	(Chandarana et al., 2021)
Impulse-regime DCQO	Keeps mainly the CD contribution in short-time evolution	(Cadavid et al., 2023)
Variationally optimized CD coefficients	Optimizes coefficients inside a fixed CD ansatz	(Sun et al., 2022)
BF-DCQO	Iterative bias-field update from measured $ZY$4	(Cadavid et al., 2024)
High-dimensional DCQC	Extends the paradigm from qubits to qutrits	(Tancara et al., 2024)

DC-QAOA is the most direct variational extension of QAOA. Its original formulation uses counterdiabatic operators derived from approximate gauge-potential theory and reports clear low-depth advantages over standard QAOA on 1D Ising models, MaxCut, SK, and $ZY$5-spin benchmarks (Chandarana et al., 2021). A later reevaluation on weighted and unweighted one-dimensional MaxCut sharpened that picture: first- and second-order CD corrections do reduce the number of QAOA steps needed for convergence, but for the model studied the total number of free parameters required to achieve the faster convergence is independent of the QAOA variant, so the benefit is a tradeoff between shallower circuits and a more complex variational landscape (Vizzuso et al., 2023).

Impulse-regime DCQO takes a more aggressive route. In portfolio optimization, a pure DCQO protocol based on the CD term alone reduced effective depth by factors of $ZY$6 to $ZY$7 relative to hybrid QAOA or digitized adiabatic baselines, with a 20-asset experiment on IonQ hardware (Cadavid et al., 2023). The same design principle reappears in logistics scheduling, where pure DCQO and hybrid DCQO were benchmarked against digitized annealing and QAOA on job-shop scheduling and TSP, with several-orders-of-magnitude success-probability improvements over QAOA at the same two-qubit-gate budget claimed in the abstract (Dalal et al., 2024).

Other work focuses on improving the counterdiabatic term itself rather than changing the overall algorithm. One study optimized the coefficients of a fixed low-order CD ansatz by a variational quantum circuit, thereby preserving the simple two-body $ZY$8 structure while improving finite-time GHZ preparation and outperforming shallow QAOA at equal bounded time (Sun et al., 2022). Another line adds meta-learning for parameter initialization in DC-QAOA using LSTM and GRU networks, with reported gains in convergence speed and final error on MaxCut and SK (Chandarana et al., 2022).

BF-DCQO departs from variational training altogether. It iteratively updates the initial Hamiltonian to

$ZY$9

so that each new run begins from a biased product state informed by prior low-energy samples (Cadavid et al., 2024). A subsequent extension to higher-order unconstrained binary optimization combined first-order CD digitization with CVaR-based low-energy-tail updates and a final weighted signed bias, and reported 156-qubit hardware experiments plus 433-qubit MPS simulations on nearest-neighbor three-local HUBO instances (Romero et al., 2024).

High-dimensional counterdiabatic quantum computing generalizes the same paradigm to qutrits. In that setting, the adiabatic-plus-CD framework is kept intact, but trinary decision variables are encoded directly into qutrit Hamiltonians for multi-way number partitioning, max 3-cut, and portfolio optimization; the reported success-probability enhancement over qubit encodings reaches factors above $P$0 for some max 3-cut instances (Tancara et al., 2024).

4. Benchmarks, problem classes, and empirical regimes

Digitized counterdiabatic quantum optimization has been tested on a wide range of problem classes, but the benchmark structure is highly uneven across the literature.

The original DC-QAOA paper emphasized 1D Ising systems, unweighted 3-regular MaxCut, the SK model, and $P$1-spin benchmarks. In one longitudinal-field Ising example on 12 qubits, DC-QAOA reached unit approximation ratio already at $P$2, whereas standard QAOA required $P$3 (Chandarana et al., 2021). For 1D weighted and unweighted MaxCut, the later convergence study confirmed faster convergence with higher-order CD corrections but also emphasized the compensating increase in cost-landscape complexity (Vizzuso et al., 2023).

Application-oriented work extended the paradigm far beyond canonical spin-glass benchmarks. A protein-folding study constructed a 5-local Ising Hamiltonian for tetrahedral-lattice folding and then attacked it with a CD-inspired variational circuit containing only one- and two-local $P$4 and $P$5 generators; the method was applied to proteins with up to 9 amino acids and 17 qubits on hardware (Chandarana et al., 2022). A logistics-scheduling study addressed laboratory robot job-shop scheduling and TSP, demonstrating both pure and hybrid DCQO on trapped-ion and superconducting devices (Dalal et al., 2024). A portfolio-optimization study used a 20-asset instance and obtained hardware results on IonQ Aria (Cadavid et al., 2023).

The benchmark scale has also expanded. BF-DCQO reported experiments on a 36-qubit weighted maximum independent set instance on trapped ions and a 100-qubit heavy-hex spin glass on IBM hardware, together with claims of polynomial scaling enhancement over traditional DCQO and finite-time AQO for all-to-all Ising spin glasses (Cadavid et al., 2024). The higher-order BF-DCQO extension ran a full 156-qubit hardware loop for nearest-neighbor three-local HUBO and weighted MAX 3-SAT, and compared against simulated annealing, Tabu search, D-Wave quantum annealing, and CVaR-QAOA on the studied instances (Romero et al., 2024). A separate experimental study on digitized counterdiabatic quantum critical dynamics used up to 156 superconducting qubits and observed up to $P$6 reduction in topological defects during fast quenches across transverse-field Ising critical points, a result presented as directly relevant to finite-time quantum optimization because reduced defects correspond to reduced residual excitation (Visuri et al., 20 Feb 2025).

These benchmarks suggest a recurrent empirical pattern. Counterdiabatic additions are most effective in shallow or fast-evolution regimes where plain digitized annealing or low-depth QAOA are limited by diabatic error. Several papers explicitly state that as depth grows, standard QAOA or standard annealing tend to catch up, while the extra counterdiabatic structure becomes less decisive (Chandarana et al., 2021). This suggests that DCQO is best understood as a low-depth enhancement strategy rather than a universal replacement for deeper variational or adiabatic schemes.

5. Hardware realization, compilation, and codesign

The practical performance of digitized counterdiabatic optimization depends heavily on compilation, routing, and hardware connectivity because the CD terms introduce noncommuting two-qubit interactions that are absent from ZZ-only QAOA.

One compilation-oriented study argued that the executable gate sequence induced by routing is part of the algorithm, not merely a backend detail, because DC-QAOA contains noncommuting $P$7 and $P$8 terms whose ordering affects both Trotter error and trainability. Its algorithm-oriented qubit mapping strategy co-optimized Hamiltonian order, variational parameters, and qubit placement, and on IBM devices reported an average approximation-ratio increase of $P$9 without error mitigation and $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$0 with error mitigation, together with average reductions of $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$1 in CX count and $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$2 in circuit depth relative to Qiskit and Tket (Ji et al., 2023).

Hardware connectivity strongly shapes which problem classes are realistic. Dense Ising and dense QUBO encodings map naturally onto trapped-ion devices with all-to-all connectivity, and several papers accordingly used IonQ or Quantinuum platforms for all-to-all portfolio, scheduling, and protein-folding circuits (Cadavid et al., 2023). Sparse architectures, especially IBM heavy-hex processors, favor nearest-neighbor or hardware-matched Hamiltonians; this is the rationale behind the 100-qubit heavy-hex BF-DCQO experiment and the 156-qubit nearest-neighbor HUBO implementation (Cadavid et al., 2024). In the critical-dynamics experiments, graph coloring was used to parallelize commuting or nonoverlapping two-qubit gates, a reminder that CD-enhanced digitized annealing is as much a scheduling problem as a Hamiltonian-design problem on NISQ hardware (Visuri et al., 20 Feb 2025).

Compression strategies are equally central. Impulse-regime DCQO drops the adiabatic terms entirely where $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$3 dominates and then prunes small-angle gates with a threshold, yielding dramatic depth reductions in portfolio optimization (Cadavid et al., 2023). Single-layer DCQO for $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$4-spin models uses a specially chosen schedule and transverse-field strength so that the useful CD action is concentrated near the midpoint, effectively collapsing the circuit to one nontrivial layer (Guan et al., 2023). In BF-DCQO, iterative biasing can eventually simplify later circuits so strongly that the final iteration of a 156-qubit weighted MAX 3-SAT experiment required no entangling gates at all, with circuit depth 4 (Romero et al., 2024).

A plausible implication is that digitized counterdiabatic optimization is unusually sensitive to algorithm–hardware codesign. Unlike standard QAOA, where the dominant issue is often the number of ZZ layers, DCQO must co-manage noncommuting Pauli strings, native-gate decomposition, layout, and schedule compression.

6. Complexity tradeoffs, controversies, and open questions

The literature is explicit that the main tradeoff is not whether CD helps, but under what cost metric it helps. Equal-depth comparisons usually favor DC-QAOA or DCQO because the added CD block makes each layer more expressive. Yet two later analyses show why this is incomplete.

For one-dimensional MaxCut, higher-order DC-QAOA converges in fewer layers, but the total number of free parameters required for the faster convergence is independent of the variant studied, implying that the gain in circuit depth is offset by a more complicated variational landscape (Vizzuso et al., 2023). For MaxCut under equal total CNOT count, DC-QAOA(NC) becomes advantageous only beyond a problem-size threshold: the reported crossover is at more than 16 qubits, because one NC counterdiabatic layer uses about three times as many CNOT gates as one QAOA layer (Liu et al., 2024). The same study also found that one-layer DC-QAOA(NC) versus three-layer QAOA, compared at equal CNOT count, shows an empirical exponential performance advantage in the reported SK benchmarks, but that claim is explicitly about the observed metric under the chosen comparison, not a worst-case complexity-theoretic separation (Liu et al., 2024).

Classical optimization remains another unresolved axis. Benchmarking work on hybrid DCQC found that optimizer choice materially changes performance and concluded that a fully parameterized DCQC ansatz paired with SPSA-based BFGS is especially effective up to 28 qubits on SK at $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$5, while PCA indicated that the first two principal components capture at least $$H_a(t) = (1-\lambda(t))H_{mixer} + \lambda(t)H_{prob},$$6 of variance in the examined landscapes (Xu et al., 2024). This suggests that trainability is not uniform across ansätze: some counterdiabatic parameterizations appear structured rather than pathological, while others induce rugged landscapes or parameter-count inflation.

Several papers also caution against overgeneralization of favorable experimental claims. The critical-dynamics study demonstrates defect suppression in clean transverse-field Ising ferromagnets, but explicitly notes that this is only indirectly evidence for broader combinatorial optimization, especially because many hard optimization instances involve first-order transitions or frustration (Visuri et al., 20 Feb 2025). The higher-order BF-DCQO paper states that its “commercial quantum advantage” language should be interpreted cautiously, because the instances are nearest-neighbor and well matched to both hardware connectivity and MPS simulation (Romero et al., 2024). Likewise, the qutrit study’s statement that it is “always better” to use high-dimensional systems is not universally true instance by instance in the supplied results, although the aggregate max 3-cut data are strongly favorable to qutrits (Tancara et al., 2024).

Open questions therefore cluster around five themes. The first is AGP quality: whether higher-order or learned CD operators can improve fidelity without losing hardware accessibility. The second is cost accounting: whether depth, parameter count, CNOT count, and wall-clock optimization cost can be combined into a single meaningful comparison framework. The third is scaling beyond structured benchmarks: many of the strongest demonstrations remain in 1D, nearest-neighbor, or specially encoded instances. The fourth is hardware-native realization: routing and native-gate synthesis materially alter the realized algorithm. The fifth is hybrid versus nonvariational control: BF-DCQO and pure DCQO avoid outer-loop optimization, whereas DC-QAOA and h-DCQO rely on it; the relative superiority appears regime-dependent rather than absolute.

Taken together, the field depicts digitized counterdiabatic quantum optimization as a low-depth optimization strategy grounded in adiabatic gauge-potential theory, but now diversified into variational, nonvariational, iterative-feedback, and high-dimensional forms. The strongest established result is not a universal asymptotic advantage claim, but a repeated and technically consistent observation: when coherence time, two-qubit-gate budget, or trainability makes conventional digital optimization shallow, counterdiabatic structure often yields a better use of the available circuit budget.